A few days ago I announced a personal challenge: reviewing AI startup documentation in public — one thing done well, one place onboarding could be smoother, one rewritten example. This is review #1.
The product: ktx by Kaelio — an open-source, self-improving context layer for data agents. It teaches coding agents like Claude Code, Codex, and Cursor how to query your warehouse accurately, using approved metric definitions and business context stored as reviewable YAML and Markdown in git. You install it with npm install -g @kaelio/ktx (Node.js 22+) and run one guided ktx setup.
I tested the docs rather than just reading them. Everything below was verified live against docs.kaelio.com on July 8, 2026.
What Kaelio does really well: docs written for agents, not just humans
This is the part I want other AI startups to copy.
Most documentation assumes a human reading a browser tab. ktx treats an AI coding agent as a first-class reader:
A dedicated AI Resources route with an Agent Quickstart, Agent Instructions, and Prompt Recipes.
/llms.txt (a curated index of high-value pages) and /llms-full.txt (the full corpus) so an assistant can discover the right pages before diving in.
Per-page Markdown on demand: curl -H "Accept: text/markdown" returns clean source — frontmatter stripped, code blocks preserved, tables preserved.
Missing pages return a plain-text 404 instead of silently falling back to rendered HTML. Small detail, big kindness — it stops an agent from confidently parsing a garbage page.
Copy as Markdown, View MD, and Copy MDX actions sit right on each rendered page.
One of my standing review questions is literally: is the documentation structured so both developers and AI agents can use it effectively? ktx is the clearest "yes" I've come across. If you're building a developer product in 2026, this is the bar.
The human path is strong too: a single guided ktx setup wizard with clearly numbered steps, a live demo warehouse with paste-ready credentials, and a Common issues table that maps symptom → fix.
Where onboarding could be smoother: the MCP step the happy path skips
Here's the one seam.
The quickstart walks you through a clean sequence: install → ktx setup → ktx status (verify) → Connect a coding agent. That final section covers installing project-local agent rules with ktx setup --agents.
What it doesn't mention is that, before your agent client can actually reach ktx, you may need to start the local MCP daemon. That step lives in the README and FAQ — not in the quickstart:
The README notes: if ktx status prints ktx mcp start --project-dir ..., run it before opening your agent client.
The FAQ explains there's no hosted service — the local MCP daemon runs on demand via ktx mcp start when an agent client needs it.
So the failure mode is: a new developer follows the quickstart end-to-end, opens Claude Code or Cursor, and the agent silently can't see ktx — with no breadcrumb in the quickstart pointing at the fix. Making it slightly stickier, the ktx status example output in the Verify section shows a fully-ready project but doesn't include the ktx mcp start line the README says status can emit — so even the example doesn't prepare you for it.
To be fair: ktx is under active development, and this is one missing signpost on an otherwise excellent path. But it sits at the last mile of onboarding — exactly where a new developer's first impression is decided.
The rewritten section
Here's how I'd close the gap — a version of "Connect a coding agent" written in ktx's own voice, folding in the missing step:
Connect a coding agent
The setup wizard installs project-local agent rules in its last step. To install or change targets later:
ktx setup --agents
Claude Code and Codex also support global installs with --global. Agent rules point at the ktx CLI path that created them, so agents don't need a separate ktx binary on PATH. If the CLI path changes, rerun ktx setup --agents.
Start the MCP server if your status asks for it. Some setups serve agents through an on-demand local MCP daemon. If ktx status prints a line like:
ktx mcp start --project-dir /home/user/analytics
run that command before opening your agent client. Otherwise the agent connects to nothing and ktx tools appear empty. You can confirm with ktx status — Agent integration should read ready.
One short subsection, one status breadcrumb. That's the whole fix — and it's the difference between a silent dead-end and a thirty-second recovery.
The pattern worth noticing
The irony here is instructive: ktx has some of the best agent-facing documentation I've reviewed, and its one gap is on the human path — a step the team likely stopped seeing because they already know it's there. That's the most common failure I find across audits: docs that are perfectly correct for someone who already understands the product, which is exactly the one audience that doesn't need them.
If you write docs, the cheapest test available is still the best one: hand your quickstart to someone (or some agent) with zero context, and watch where they stall.
Next review
If you're building an AI product with public documentation, I'd love to include it — drop a link in the comments or reach out. Same format every time: one thing done well, one gap, one rewrite.
I'm Virginia Mwega — a full-stack & AI engineer specializing in documentation engineering and developer experience. Portfolio: virginia-mwega.vercel.app · Writing: virginiamwega-com.vercel.app · Connect: https://www.linkedin.com/in/virginia-mwega-196309313/
Top comments (0)